Automated recognition of Meso-Cenozoic foraminifera from Senegalese sedimentary deposits using convolutional neural...

Fossil foraminifera are key proxies for biostratigraphy and paleoenvironmental reconstructions, providing crucial insights into past ocean conditions and climate evolution. However, their identification remains time-consuming and challenging, particularly in taxonomically complex assemblages and in samples affected by diagenetic alteration. Here, we investigate the application of convolutional neural networks (CNNs) for the automated classification of Meso-Cenozoic foraminifera from the West African margin, including planktonic, benthic, and agglutinated taxa. Three training datasets comprising 11,378 images were used to develop CNN models for microfossil detection, genus-level, and species-level classification. Genus-level classification reached a validation accuracy of 72.4%, outperforming species-level identification (50.2% validation accuracy), as it reduced misclassifications from morphological variability and post-mortem degradation in closely related species. The genus-level CNN performed well for genera with distinct morphological features, such as Muricohedbergella sp. and Heterohelix sp., but struggled with Subbotina spp. and Dicarinella spp. due to shared morphological traits and diagenetic alterations. For benthic foraminifera, Nummulites sp. was easily identified, while Gavelinella sp. posed significant challenges. These results highlight the potential of CNNs for high-throughput classification, while also revealing limitations particularly for taxa with high morphological variability or diagenetic changes. The findings have important implications for biostratigraphy, suggesting that CNNs can enhance genus-level classification in biostratigraphic applications. Further advancements such as multi-view imaging and expanded training datasets could enhance the CNN performance. By integrating deep-learning-based classification with expert biostratigraphy, our study contributes to open the way for quantitative and reproducible reconstructions of Meso-Cenozoic stratigraphic constraints.